Dharma Field Extractor: Any Document In, Structured Fields in JSON Out
Reported by Hugging Face and ecosistemastartup2 sources ↓
Dharma-AI introduced a 2-billion-parameter model that extracts requested fields from document images into JSON. The company reports a 0.834 benchmark score and estimated running costs below those of the frontier-model configurations it tested.
Key points
- The model returns requested document fields as typed JSON, using an open request or a specified schema
- Dharma-AI reports a 0.834 field-level F1 score across 1,638 benchmark cases
- The company says estimated costs were about 6 to 147 times lower than the tested frontier models
- The model is based on DharmaOCR-LITE v1.1 and a LoRA adapter trained on 43,384 examples
Sources
Details
Dharma-AI says the model can extract fields from document images either by identifying available information or by filling a supplied schema. It returns JSON, with null for requested fields that are absent; schema-based output uses the specified keys without extra fields. The company says the same model handles different document types without per-type fine-tuning.
The company’s benchmark covered 546 documents in three request scenarios, for 1,638 cases. Its field-level F1 metric accounts for field paths, value similarity and whether the output is valid JSON. Dharma-AI reports a score of 0.834, compared with 0.816 for Claude Opus 5 and 0.741 for GPT-5.6 Luna; the benchmark was published by the company and has not been independently validated in the articles.
Dharma-AI estimates costs were roughly six times lower than GPT-5.6 Luna and 147 times lower than GPT-5.6 Sol. Those estimates compare measured token use and providers’ published prices for frontier models with GPU time for the self-hosted model, running on one Nvidia L40S at $2.24 per hour. The company says actual deployment costs depend on hardware, volume and latency requirements.
The model uses DharmaOCR-LITE v1.1, a 2-billion-parameter OCR model, with a LoRA adapter trained on 43,384 examples spanning receipts, invoices, forms, bank statements, IDs and synthetic documents. Dharma-AI also reports results on SROIE and FUNSD, public datasets it says were excluded from training.
Written from coverage as of
Images
- Hugging Face
- ecosistemastartup
- Hugging Face
- Hugging Face
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh